Can Artificial Intelligence Detect the Earliest Signs of Heart Disease? What a Major CT Angiography Study Reveals About Tiny Coronary Plaques

Table of Contents

• Key Points
• Why This Research Matters: The Problem of Early Atherosclerosis
• Study Methods: How the Research Was Conducted
• Key Findings: What Happened to the Small Plaques
• The Role of Statins in Plaque Changes
• Clinical Implications: What This Means for Patients
• Study Limitations: What This Research Couldn't Prove
• Recommendations: Actionable Advice for Patients
• Frequently Asked Questions
• Source Information

Key Points

• In a 99-patient study, AI-QCT detected tiny coronary plaques that human readers often miss.
• 87% of small plaques persisted at the same location on follow-up scans after about 3.8 years.
• Median plaque volume tripled from 7.1 mm³ to 18.9 mm³ in persistent plaques.
• Statins increased plaque calcification, potentially stabilizing soft plaque, but didn't reduce total volume.
• No clinical outcomes were tracked; larger studies are needed to confirm findings.

Why This Research Matters: The Problem of Early Atherosclerosis

Atherosclerosis — the buildup of fatty plaque inside artery walls — remains the leading cause of cardiovascular death worldwide. Coronary artery disease (CAD) alone affects nearly 200 million people globally and causes more than 9 million deaths each year. It also contributes to approximately 182 million disability-adjusted life years, a measure of healthy life lost to disease and disability.

The news isn't all bad. Prior declines in heart disease deaths had been a public health success story. But those gains are now being threatened by rising rates of obesity and diabetes, which fuel the development of atherosclerosis.

One of the most important discoveries in recent cardiology is that you don't need a severely blocked artery to be at risk. Studies have convincingly shown that non-obstructive plaque — plaque that narrows the artery by less than 50% — plays a major role in causing heart attacks and cardiovascular death. On average, patients with non-obstructive CAD have an 8-fold higher annual event rate compared with patients who have no coronary atherosclerosis at all. Even after adjusting for other risk factors, the hazard ratio for major adverse cardiovascular events ranges from 1.5 to 7.2 when comparing patients with non-obstructive CAD to those with no CAD. This means such patients are anywhere from 50% to over 7 times more likely to experience a heart attack, stroke, or cardiac death.

But a critical question has remained unanswered: do these risks also apply to extremely small plaques that are barely visible on imaging? And is there a volume threshold below which risk hasn't yet begun to climb?

There has been a major obstacle to answering these questions. Small, non-calcified plaques are notoriously difficult for human readers to spot on coronary CT angiography (CCTA). They can look almost identical to the soft tissue surrounding the coronary arteries, to other soft tissue structures, or simply to image noise and artifacts. Because of this uncertainty, such tiny plaques are frequently not reported at all.

This is where artificial intelligence enters the picture. A new technology called atherosclerosis imaging quantitative computed tomography (AI-QCT) uses machine learning to automatically detect, measure, and characterize plaque throughout the entire coronary tree. In clinical practice, AI-QCT frequently identifies small plaques that human readers miss. But before AI-detected tiny plaques can be used to guide patient care, researchers needed to verify that these detections represent real atherosclerosis and not just software artifacts.

Study Methods: How the Research Was Conducted

The researchers turned to the PARADIGM study (Progression of Atherosclerotic Plaque Determined by Computed Tomographic Angiography Imaging), a large international registry that enrolled 2,252 patients from 7 countries and 13 different sites. All participants underwent serial (repeat) CCTA scans for known or suspected coronary artery disease. Patients were enrolled between 2003 and 2015.

For this specific analysis, the research team restricted inclusion to the first 99 patients who underwent CCTA in the cohort. This number was deliberately chosen as a feasibility study, without formal statistical power analysis for any particular outcome. An additional 2 patients were screened but excluded because they did not have any qualifying small plaques.

Each patient underwent a baseline scan (CCTA-1) and a follow-up scan (CCTA-2) at least 2 years apart. The scans were performed according to Society of Cardiovascular Computed Tomography guidelines, using 64-detector row or newer scanners with either single- or dual-source technology.

The key innovation was in how the scans were analyzed. A commercially available, FDA-cleared AI software (Cleerly Lab, Cleerly, Denver, CO) performed automated analysis using validated convolutional neural network models. The AI system automatically:

• Segmented the coronary arteries
• Identified vessel and lumen contours
• Labeled each coronary segment
• Calculated the percentage of diameter stenosis
• Characterized plaques by type

Plaques were classified using Hounsfield unit (HU) density measurements, a scale that quantifies how dense a tissue appears on CT:

• Calcified plaque: density greater than 350 HU
• Non-calcified plaque: density between 30 and 350 HU
• Low-density non-calcified plaque: density less than 30 HU

Importantly, all AI output was verified by a Level III experienced reader — the highest level of CCTA certification — who was blinded to whether each scan was the baseline or follow-up study. This blinding was a critical design element to prevent bias.

Small plaques were defined as those with a total plaque volume of 0.1 to 50 mm³. For context, 50 mm³ is roughly the size of a small grain of rice.

The researchers examined three specific questions:

1. What proportion of small plaques identified on the baseline scan were also present at the same location on the follow-up scan? This provided a measure of the software's reliability — plaques that persisted at the same location likely represent true atherosclerosis, while those that vanished could represent either true regression or false-positive detection on the first scan.
2. What are the characteristics of small plaques, including their calcified, non-calcified, and low-density non-calcified components?
3. What features are associated with progression of small plaques over time?

Statistical analysis was thorough. The team used Student's t-tests and Mann-Whitney U tests for continuous data, Wilcoxon signed-rank tests for paired CCTA characteristics, and Chi-square or Fisher exact tests for categorical data. Logistic regression was used to model which plaques disappeared versus persisted, and linear regression was used to evaluate the impact of statin therapy on plaque volume changes. All tests were two-tailed, with statistical significance defined as p-values below 0.05.

Key Findings: What Happened to the Small Plaques

A total of 99 patients with 502 small plaques were included in the analysis. The participant characteristics paint a picture of a typical middle-aged, at-risk population:

• Median age: 61 years (interquartile range 54–67)
• Male gender: 63%
• Hypertension: 55.6%
• Current or prior smoking: 40.4%
• Family history of coronary artery disease: 33.4%
• Hypercholesterolemia (high cholesterol): 31.3%
• Diabetes: 12.1%
• Body mass index (BMI): median 25.2 kg/m² (23.9–27.6)
• Statin use at baseline: 44.4%

At the time of the baseline scan, the median total plaque volume was 6.8 mm³ (interquartile range 3.5–13.9 mm³). The overwhelming majority of this was non-calcified plaque, with a median non-calcified volume of 6.2 mm³ (2.9–12.3 mm³). Calcified plaque volume was essentially zero in most patients.

The mean time between the baseline and follow-up scans was 3.8 ± 1.6 years. What happened to these tiny plaques over that period is the central finding of the study:

Finding #1: 87% of small plaques persisted at the same location.

On follow-up imaging, 437 of 502 plaques (87%) were found at the exact same location as the baseline scan. Among those persistent plaques, 72% had grown larger, while 15% had decreased in volume.

Finding #2: Small plaque volume tripled over roughly 4 years.

For plaques that persisted, the median total plaque volume increased dramatically from 7.1 mm³ (IQR 3.8–15.4 mm³) at baseline to 18.9 mm³ (IQR 8.3–45.2 mm³) on the follow-up scan. That is nearly a 3-fold increase. The non-calcified component grew from 6.7 mm³ to 13.8 mm³, and calcified plaque volume increased from 0 mm³ to 2.5 mm³. The diameter stenosis (the percentage of the artery blocked) also increased from a median of 6% to 13%.

Finding #3: Smaller plaques were less likely to persist.

Plaques with a volume under 2 mm³ at baseline persisted in only 41 of 62 cases (66%). In contrast, plaques larger than 2 mm³ at baseline persisted in 395 of 439 cases (90%). This suggests that extremely tiny plaques are more difficult to distinguish from image noise, though even among these, two-thirds were confirmed to be real atherosclerosis.

Finding #4: Plaques that "disappeared" had distinct characteristics.

The 65 plaques (13%) that were seen on the baseline scan but not on follow-up were significantly different from those that persisted (all p-values < 0.05):

• Lower total plaque volume: 3.9 mm³ vs. 7.0 mm³
• Shorter plaque length: 4.5 mm vs. 6.0 mm
• More distal location (further away from the artery opening): 21.8 mm vs. 12.6 mm from the ostium
• Less likely to have a calcified component
• Less severe diameter stenosis: 3% vs. 6%

Interestingly, none of the patient-related factors — age, sex, hypertension, smoking history, BMI, family history, hypercholesterolemia, diabetes, or baseline statin use — predicted whether a plaque would persist or disappear. Likewise, scanner-related parameters (vendor, tube current, tube voltage) made no difference. In the multivariable model, only smaller total plaque volume remained independently predictive of a plaque being absent on follow-up.

The Role of Statins in Plaque Changes

Statin use increased significantly during the study period. At baseline, 48% of patients were taking statins. By the follow-up scan, that number had risen to 68% (p < 0.001).

Among 429 small plaques with available statin data, 145 plaques (33.8%) occurred in patients who were still not taking statins at the time of the follow-up CCTA.

The researchers compared plaque volumes between statin users and non-users at follow-up:

• Total plaque volume: 19.4 mm³ in statin users vs. 18.15 mm³ in non-users — not statistically significant (p = 0.21)
• Non-calcified plaque volume: 12.8 mm³ vs. 14.35 mm³ — not significant (p = 0.68)
• Calcified plaque volume: 3.9 mm³ in statin users vs. 0.2 mm³ in non-users — highly significant (p < 0.001)

This finding aligns with a well-established biological effect of statins: they stabilize plaque by promoting calcification. A growing body of research shows that statin therapy increases the calcium content of atherosclerotic plaque, which paradoxically makes plaque more calcified but less dangerous. The fact that statin users showed significantly more calcified plaque volume supports the conclusion that the AI-QCT findings represent true biological atherosclerosis rather than imaging artifacts.

In the regression analysis, statin use at follow-up was significantly associated with changes in calcified plaque volume (p < 0.001) but not with changes in total plaque volume (p = 0.30), non-calcified plaque volume (p = 0.70), or low-density non-calcified plaque volume (p = 0.68).

Clinical Implications: What This Means for Patients

This study carries several important messages for patients and their doctors.

First, AI can detect real atherosclerosis at remarkably early stages. The fact that 87% of small plaques identified by AI-QCT were confirmed at the same location years later — and often grew substantially — demonstrates that these tiny findings represent genuine disease, not software errors. For patients, this means AI-enhanced CT scans may be able to identify heart disease years or even decades before it would cause symptoms.

Second, very small plaques are not "harmless" — they tend to grow. The finding that median plaque volume tripled over an average of 3.8 years is a powerful reminder that atherosclerosis is a progressive disease. A plaque that starts as a 7 mm³ speck can become a 19 mm³ plaque within a few years. While this study did not track clinical outcomes, the broader literature shows that total plaque burden is the main independent predictor of major adverse cardiovascular events.

Third, statins appear to change the composition of plaque, even when they don't shrink it. The significantly higher calcified plaque volume in statin users suggests that these medications help convert dangerous soft plaque into more stable calcified plaque. This is consistent with how cardiologists understand statins: they stabilize the "ticking time bomb" plaques that are most likely to rupture and cause heart attacks.

The study also highlights a broader trend in preventive cardiology. Prior research in a cohort of nearly 24,000 symptomatic patients referred for cardiac CT found that obstructive coronary artery disease was not associated with higher risk than non-obstructive plaque when patients were stratified by total calcified plaque burden. In other words, it's not just about whether an artery is blocked — it's about how much plaque you have throughout your entire coronary tree. Recent data also suggest that even small plaque volumes (>0–250 mm³) increase the 10-year incidence of major adverse cardiovascular events compared with no plaque at all.

Study Limitations: What This Research Couldn't Prove

It's important to understand what this study could not show.

The sample size was small. Only 99 patients were included, and this was explicitly designed as a feasibility study without formal statistical power calculation for any specific outcome. Larger studies will be needed to confirm these findings.

No clinical outcomes were tracked. The study measured plaque progression on imaging, not heart attacks, strokes, or deaths. While plaque growth is a well-established surrogate marker for risk, the study does not directly prove that AI-detected small plaques predict future clinical events.

The "gold standard" question remains. There is currently no non-invasive gold standard for distinguishing true small plaques from imaging artifacts. The researchers used follow-up CCTA as their reference — reasoning that a plaque present at the same location years later must have been real at baseline. This is a clever approach, but it cannot fully distinguish between true plaque regression and false-positive detection on the baseline scan for the 13% of plaques that disappeared.

Human reader variability is a known challenge. Previous research cited in this paper found that the kappa coefficient — a statistical measure of reader agreement — was only 0.52 (95% CI 0.49–0.55) for intraobserver agreement and 0.46 (95% CI 0.43–0.49) for interobserver agreement when classifying CCTA as no disease, mild, moderate, or severe. The authors note that agreement would likely be even lower for distinguishing small plaques from no plaque, which underscores the potential value of AI assistance.

The study population may not represent all patients. Participants were enrolled from 7 countries but were predominantly middle-aged, and the gender split was 63% male. Results may differ in younger patients, women, or more diverse populations.

Recommendations: Actionable Advice for Patients

Based on this research and the broader medical literature, here are practical steps patients can consider:

1. Know your plaque burden, not just your cholesterol numbers. If you have risk factors for heart disease, talk to your doctor about whether a coronary CT angiography or a coronary calcium score is appropriate for you. The total amount of plaque in your coronary arteries is a powerful predictor of risk.
2. Don't dismiss "mild" findings. If a CT scan shows non-obstructive plaque or even very small plaques, this is not a clean bill of health. The data are clear: even non-obstructive plaque raises your risk of heart attack compared to having no plaque at all.
3. Consider statin therapy if your doctor recommends it. This study reinforces that statins change plaque composition in favorable ways — increasing calcium content and stabilizing dangerous soft plaque — even when they don't dramatically shrink overall plaque volume. The decision to start a statin should be individualized, but patients with documented coronary plaque are often excellent candidates.
4. Address all modifiable risk factors. In this study, 55.6% of participants had hypertension, 40.4% had a smoking history, and 31.3% had high cholesterol. Controlling blood pressure, quitting smoking, managing cholesterol, maintaining a healthy weight, and controlling diabetes remain the cornerstones of preventing atherosclerosis progression.
5. Ask about AI-enhanced imaging. If you are undergoing cardiac CT, ask whether AI-based plaque analysis is available. This study suggests that AI software can detect very small plaques that human readers may miss, providing a more complete picture of your cardiovascular risk.
6. Remember that early detection opens the door to prevention. The fact that small plaques can be reliably detected — and that they grow over time — means there is a window of opportunity to intervene with lifestyle changes and medication before plaque becomes extensive.

The takeaway message is hopeful. The ability to detect atherosclerosis at its earliest stages, when a plaque is just a few cubic millimeters in size, represents a genuine advance in preventive cardiology. With tools like AI-QCT, doctors may soon be able to identify patients at risk years earlier than previously possible — and intervene with statins, lifestyle changes, and risk factor control while the disease is still in its infancy.

Frequently Asked Questions

What is AI-QCT and how does it detect small coronary plaques?

AI-QCT (atherosclerosis imaging quantitative computed tomography) is an FDA-cleared artificial intelligence software that automatically detects, measures, and characterizes plaque throughout the coronary arteries. In a feasibility study of 99 patients, it identified tiny plaques as small as a fraction of a cubic millimeter that human readers often miss. All results were verified by a Level III reader.

Are very small coronary plaques dangerous or likely to grow?

In a study of 99 patients followed for about 3.8 years, 87% of small plaques persisted at the same location, and among persistent plaques, 72% grew larger. Median plaque volume tripled from 7.1 mm³ to 18.9 mm³. This shows that even extremely small plaques tend to progress over time.

Should I ask my doctor for an AI-enhanced coronary CT scan?

If you have risk factors for heart disease, talk to your doctor about whether a coronary CT or calcium score is appropriate. This study suggests AI-based analysis can detect very small plaques that human readers may miss, providing a more complete picture of risk. However, the decision should be individualized.

What are the limitations of this study on AI plaque detection?

This was a small feasibility study with only 99 patients, and no clinical outcomes such as heart attacks were tracked. The reference standard was follow-up CT, not a gold standard. Results may differ in younger patients, women, or more diverse populations, so larger studies are needed to confirm findings.

Source Information

This patient-friendly article is based on peer-reviewed research published in the Journal of Cardiovascular Computed Tomography, volume 17 (2023), pages 407–412.

Original article title: How early can atherosclerosis be detected by coronary CT angiography - Cleerly James Min

Authors: Rhanderson Cardoso, Andrew D. Choi, Arthur Shiyovich, Stephanie A. Besser, James K. Min, James Earls, Daniele Andreini, Mouaz H. Al-Mallah, Matthew J. Budoff, Filippo Cademartiri, Kavitha Chinnaiyan, Jung Hyun Choi, Eun Ju Chun, Edoardo Conte, Ilan Gottlieb, Martin Hadamitzky, Yong-Jin Kim, Byoung Kwon Lee, Jonathon A. Leipsic, Erica Maffei, Hugo Marques, Pedro de Araújo Gonçalves, Gianluca Pontone, Sang-Eun Lee, Ji Min Sung, Renu Virmani, Habib Samady, Fay Y. Lin, Peter H. Stone, Daniel S. Berman, Jagat Narula, Leslee J. Shaw, Jeroen J. Bax, Hyuk-Jae Chang, and Ron Blankstein

DOI: 10.1016/j.jcct.2023.08.012

Funding and disclosures: Several authors are affiliated with Cleerly Inc., the company that manufactures the AI-QCT software used in this study. The research was an open-access article published under the CC BY-NC-ND license. This patient summary was created by a medical writer and is intended for educational purposes only. It does not constitute medical advice. Patients should discuss their individual cardiovascular risk and treatment options with their healthcare provider.